Open-sourcing AstaBrief, the fast report-generation model in Asta

Allen Institute for AI open-sourced AstaBrief 8B, the Fast-mode report generator in its Asta scientific agent platform, alongside training data and an example PDF-to-report workflow. Built from Qwen3-8B with supervised fine-tuning and direct preference optimization — not RL — it turns a research question plus retrieved excerpts into a cited report in one pass.

Across Asta’s pipeline, Fast mode averages 51.1 seconds per report versus 178.5 seconds for Claude-powered Thinking mode (~3.5× faster). Training used filtered real Asta queries (~90K research-focused after cleanup), ~47K SFT examples from ScholarQA pipelines backed by then-frontier models, and ~6K DPO pairs with dual-judge agreement. Ai2 notes most training and evaluation finished in 2025 and was not fully rerun against today’s frontier.

This is the kind of useful AI we want more of: domain-specific, open weights, fast enough to iterate, and honest about what the evals actually measured. Scientists need grounded synthesis they can verify — not a polished paragraph that quietly widens a study’s claims. Ai2’s clearest lesson is almost anti-climactic: filtering synthetic reports for low citation density beat elaborate filter combos.

I’m cheering the methodological humility. They considered RL-with-judges (à la their DR Tulu line), then chose cheaper SFT+DPO so they could debug. Dual judges (GPT-4.1 and DeepSeek-R1) with claimed 95% human agreement on preferences is a practical noise cut. Early usage signals among 374 Fast-mode users — 29.1% multi-day return, 23% never switching back to Thinking — suggest speed is a real product feature for literature work.

Name the limits. Metrics emphasize rubric coverage, answer precision, and citation precision/recall — not whether the model preserves evidentiary scope (sample findings turned into population claims, past-tense results into eternal truths). Ai2 flags that gap themselves. Distillation from proprietary Claude/o-series pipelines also means “open” here is open weights and recipe, not a from-scratch scientific stack.

Our read: institutions that need air-gapped literature synthesis should try this. The win isn’t beating Claude forever — it’s proving specialized post-training and attribution filters can deliver usable scientific reports without renting a frontier API for every draft.

AstaBrief sits in Ai2’s broader science-model line (ScholarQA, DR Tulu, NSF OMAI) and ships inside Asta beside a slower Thinking mode for heavier jobs.